Seminar by Roger Kingdon

Anytime planning for advanced agent behaviour

Roger Kingdon
Dstl, Oxford, United Kingdom
Date: Monday, October 26, 2009
Time: 4:00 PM
Venue: CS102.

Abstract:

This talk reassesses the anytime approach to time-dependent planning by setting it within the broader context of contingency planning. The significant contributions of Zilberstein and Hawes to the anytime approach are described in some detail. Two practical difficulties are encountered: first, there are remarkably few demonstrated applications; and, secondly, none of the applications which have been demonstrated appear to be particularly dependent upon their anytime planning solutions. Taking the broader view, it is observed that a number of contingency planning problems that are known to be challenging to humans have several key features in common with typical multi-agent system applications. Accordingly we can compile a list of criteria for assessing the suitability of specific, well-defined problems as candidates for contingency or anytime planning. This approach establishes the Slam card game as a potential candidate, an assessment that is strengthened when it becomes clear that it is relatively straightforward to define reactive, planning and anytime agents to play this game. A program has been written which plays autonomous computing agents against one another in the game of Slam. It is found that the 'anytime' agent which comprises a simple meta-level switch between reactive and planning modes of play outperforms the other agents, regardless of how they perform in relation to one another. This result is taken to confirm the earlier analysis that placed the anytime approach within the general domain of contingency planning.
In preparation for this seminar the attendees might like to acquiant themselves with the rules of the Slam card game (search for this phrase on the internet). It's quite fun.

About The Speaker:

Roger Kingdon is a computational physicist with over 20 years of postdoctoral experience in the theory and modelling of a wide range of physical systems using both conventional and novel techniques. He has 28 publications in the open literature and he is a named author on 47 other unclassified reports. In addition he has have maintained a strong spare-time interest in artificial intelligence which has resulted in the development of a novel cognitive architecture known as 'IDEAL', and a book, 'Principia Intellegentia: The principles governing human and machine intelligence' (Allied Publishers 2009). Roger has just completed a one-year MSc in Advanced Computing at Imperial College, London, and is visiting India for 3 months prior to returning to his post in UK government research.

Back to Seminars in 2009-10